Explaining Relation Classification Models with Semantic Extents

In recent years, the development of large pretrained language models, such as BERT and GPT, significantly improved information extraction systems on various tasks, including relation classification. State-of-the-art are highly accurate scientific benchmarks. A lack explainability is currently a complicating factor in many real-world applications. Comprehensible necessary to prevent biased, counterintuitive, or harmful decisions. We introduce semantic extents, concept analyze decision patterns for classification task. Semantic extents most influential parts texts concerning Our definition allows similar procedures determine humans models. provide an annotation tool software framework models conveniently reproducibly. Comparing both reveals that tend learn shortcut from data. These hard detect with current interpretability methods, input reductions. approach can help eliminate spurious during model development. increase reliability security natural processing systems. essential step enabling applications critical areas like healthcare finance. Moreover, our work opens new research directions developing methods explain deep learning